Lattice Semiconductor, known for low-power FPGAs, is on a steep growth curve. Q3 2026 revenue was $291 million**, up **118% year-over-year** to a **record high**; first-three-quarter revenue was **$663 million, up 75.65%. With AMD (Xilinx) and Intel (Altera) holding about 84% global share and domestic vendors catching up via price hikes and longer lead times, Lattice is betting on low power and physical AI.
Xu Honglai, president of Lattice's Asia-Pacific region, said at the 9th CIIE that data centers bring much growth, but he focuses more on physical AI as the future growth point. Lattice showcased the Mach-N2 secure control FPGA series for modern infrastructure, the AI-driven Lattice Prompt design tool, and the latest Lattice Sentry solutions, presenting a security control and intelligent management ecosystem from data centers to edge.
FPGAs are called universal chips, but development barriers are high—RTL coding, timing optimization, and on-board debugging each need senior engineers. Lattice Prompt aims to lower these barriers. Launched in September, version 1.0 initially connected the Radiant tool with large language models. Xu said 1.0 was the first step; recent updates embed Lattice's accumulated knowledge and capabilities into the tool. He noted the key is writing natural-language requirements clearly so the tool understands accurately; customer habits are changing, and accurate requirement expression is key to real AI-assisted FPGA development. The tool helps with RTL coding and timing optimization, and assists code deployment and issue localization after designs reach customer boards. On compute and cost from LLM connections, Xu sees limited impact—models are pre-trained, customers need not train them, the tool can connect different LLMs for cost-appropriate choices, and the efficiency and time-to-market value far exceeds model usage costs.
Xu sees FPGAs' fundamental advantages as low power, flexible programmability, and low latency. Data centers are a key growth source: domestic cloud providers, server makers, and ODMs widely use Lattice FPGAs in control chains—power-on and thermal management for CPUs, GPUs, and board chips, and liquid-cooling leak detection. As servers move from single boards to complex racks, control tasks increase; on a leak, FPGAs must quickly decide which devices to power down and which valves to close, and fast response suits such control. Switch boards connecting many GPUs also add management and control needs. In May 2026, Lattice announced a $1.65 billion acquisition of AMI, a cloud and AI platform firmware and infrastructure management player, combining its platform firmware and management with FPGAs to expand from board-level control to full rack management. Xu said single-device prices may not be high, but overall value grows with volume.
Beyond data centers, post-quantum security is another focus. As quantum computing advances, current encryption faces challenges, and key industries now require quantum-attack resistance in equipment procurement; Lattice brought related solutions to secure control.
Physical AI is the future Xu repeatedly emphasized. He sees edge AI best understood within physical AI. In robots, cameras and sensors can feed FPGAs for aggregation and preprocessing, then connect via Ethernet or PCIe to GPUs—reducing GPU preprocessing compute, stabilizing data transmission, and improving decision timeliness. Lattice is in mass production and working with domestic and international robot customers. In medical scenarios, Lattice uses FPGAs directly for AI processing; at CIIE it demonstrated a super-resolution solution—some endoscope cameras are still 720p while displays are 4K, and Lattice uses AI to boost resolution. Xu splits physical AI into decision, perception, and action, all involving functional safety. For humanoid robot joints—sealed, fanless, small systems—motor control needs low power and small size; integrating DSP and MCU work into an FPGA can shrink size while keeping performance and response, and Lattice has provided solutions to domestic humanoid robot makers. But he does not expect physical AI to explode like LLMs—it must solve physical model problems and make simulation training and real environments more consistent.
On localization, Xu said Lattice began supply chain years ago, developing China-fit packaging with local partners, but its biggest internal investment is local application engineering—more engineers, equipment, and resources. CIIE is a window to broader customers; Lattice has exhibited via the Lin-gang platform for three consecutive years. On competition, Xu said Lattice faces not only other FPGA vendors but other technology routes; its strategy is summarized as three S's—Supply, Service/Support, and Solution—with supply as the base and local support and pre-built solutions more important.
On China's semiconductor industry, Xu said the market is large with growth potential, capital entry is normal, but chip manufacturing, equipment, and materials need long-term accumulation. He stressed inclusiveness and open cooperation—no country can complete the entire semiconductor supply chain alone. On memory shortages and price hikes, Xu said a small portion of Lattice products are packaged with memory and may be affected, but volume is small; the bigger impact is on customers' complete systems—data center customers need CPUs, HBM, and DDR together, and missing one can delay system procurement and delivery. Lattice cannot control DDR capacity, but will emphasize interface compatibility in new products and adjust controllers and IP based on available devices.
ICgoodFind Summary: Lattice uses AI tools as the entry point to lower FPGA development barriers; Lattice Prompt connects natural-language requirements directly to RTL coding and timing optimization, targeting FPGA's long reliance on senior engineers. Data center control chains and physical AI each have landing points—the former raises per-system value via rack-level management, the latter enters via robot sensor preprocessing and joint control—but physical AI will not explode as fast as LLMs, and simulation-real environment consistency remains a prerequisite.
